Recommender System(CSE4077)

Fashion Recommender System(J Component)

By:

Lokesh Kanna 19MIA1014
Nithya Sharma 19MIA1028
Yuvashree R 19MIA1053
Niharika Samyuktha K 19MIA1083

Importing dataset from kaggle (https://www.kaggle.com/paramaggarwal/fashion-product-images-small)

Pre-processing

Merging the csv data with image data

Displaying count of every category of wearable/Clothing products

Model Building

Basically the input image and fashion dataset are put in the transfer learning models (Densenet161,vgg16,resnet50) as the base of the network and compared and once the features are learned from one problem it is then used to leverage them on a new similar problem.

VGG Model

DenseNet Model

Resnet Model

Test Image

Model Prediction

We use Cosine similarity algorithm to find similarity. In general, Cosine similarity measures the similarity between two vectors of an inner product space

Visualization of Latent Space of Contents

Latent Space is a lower dimensional manifold of the high dimensional images where we expect all the instances of the dataset to lie in proximity.

Since our latent space is not two-dimensional, we will use TSNE to reduce dimensionality, so we can use some visualizations to analyse the spread of the dataset. One is to look at the neighborhoods of different classes in the latent 2D plane You can see different clusters generated by embeddings, which reinforces that features make sense. In the proceeding plots of different transfer learning model embbeding, it is possible to observe the separation by Category, and in more detail by subcategory.

Visual Analysis of the results from Vgg16 model

Here we have done clustering-based classification chart for the Master Categories: “Apparel”, “Accessories”,” Footwear”, “Personal Care” and “Free Items”. From the visualization it is clearly depicted that clothes and footwear have the most unique value and more collection of products to recommend. Vgg16 has provided least relevant results compared to other models

Visual Analysis of the results from Densenet121 model

This model is so far the best pre trained model for recommending all categories of product. From the results the top 5 recommendation is noted to be spot on for every test input and the model is accurate and consistent.

Visual Analysis of the results from Resnet50 model

This model is near accurate and provides better clustering results than Vgg16 model but both Resnet and Vgg16 were not constant over every other product category. i.e., the result in Vgg16 was 40 to 50 percent accurate, densenet121 marked a 100 percent accuracy rate and Resnet50 provides 90 percent accurate results.

As we can see in the visualization of the latent space that the similar object have formed clusters while different objects are farther from each other in the latent space. For example, various types of topwears, watches, shoes or bags form respective clusters. While visually dissimilar objects e.g. topwears vs shoes have the highest distance in latent space. One thing is to be noted that since t-SNE embedding is stochastic, the results may appear slightly different every time it is re-run. we visualized the latent space using t-SNE embedding. Then we embedded the data into Latent Space and visualized the results